Coherence-based Computational Agency I Setting the Stage 1

In this book we address the problem of introducing flexibility and adaptability in autonomous agent design in the context of agents situated in regulated environments. We argue that current cognitive architectures such as those based on BDI theory fall short in performing autonomous reasoning in agents. One of the important reasons is the lack of clear motivations for choosing a goal or an action to pursue. Instead of the intention-driven philosophy in a BDI architecture, we need a formalism which would dynamically select goals and intentions considering constraints among cognitive elements. Hence, our central thesis in this book is that agent architectures need to incorporate a motivational criterion which can be computed in terms of their cognitive elements while preserving those formal properties that make BDI-based architectures attractive. This book proposes the cognitive theory of coherence as one such motivational criterion for agents to reason and take autonomous decisions. The cognitive coherence theory we use is the one proposed by Paul Thagard. The term coherence is defined as the quality or the state of cohering, especially a logical, orderly, and aesthetically consistent relationship of parts. A coherent set is interdependent such that every piece of information in it contributes to the coherence. We take Thagard’s proposal of coherence as that of maximising satisfaction of constraints between pieces of information and put to use in the design of autonomous agents. This book advances the state of the art by proposing a computational formalisation of coherence that includes a mechanism to compute coherence values between pairs of pieces of information by formalising deductive coherence. A central contribution in this book is the proposal of a coherence-based agent architecture which extends a BDI architecture with the notion of coherence. This architecture while preserving those formal properties of BDI-based architectures, incorporates coherence as the central motivational drive and reasons under uncertainty. Based on experimental evaluation, we prove the feasibility of coherence-driven agents and show that their performance match that of humans. We further extend coherence-based architecture for normative agents to reason about norms and interact in a normative environment. Thus, we show that coherence-driven normative agents can be put to use in the evolution of the behaviour of agents and of the contents of regulatory systems. Finally, coherence as a motivational criterion is contrasted against other forms of motivational theories of agency.